Visual Saliency Detection Based On Convolutional Features Fusion

In the era of big data,the amount of imformation contained in image resources has maintained rapid growth.For the reason that the technology of refining the effective information of images has important pratical value.Visual saliency detection aims to detect the most distinctive salient objects and screen out the complete target content,which helps ease the difficulty of complex visual tasks.As the basic preprocessing steps,the development of deep convolutional neural networks has brought a qualitative leap to the task of visual saliency detection.How to build an efficient network model,fuse convolutional features and fully tap the saliency cues has gradually become the dominant directions of visual saliency detection research.The article achieve visual saliency detection based on convolutional features fusion.Specific studies are as follows:(1)To tackle the problem of object imperfection and region abruption in existing deep visual saliency detection algorithms,a multiscale context enhanced fully convolutional network base on non-local deep features is proposed.The network includes four modules which are resposible for muti-level features extraction,multiscale context feature enhancement,contrast feature extraction and local-global fusion.Firstly,multi-level local features are extracted from pretrained VGG16 model.Secondly,multiscale context information is exploited to enhance the local features.Then,combined loss function is designed to extract contrast features.Finally,the saliency map is predicted in the local-global fusion way.Expermental results on different datasets show that the proposed method has more effective suppression on background noise,and the result salient object regions are more consistent and complete.(2)In order to ease the decline in the quality of visual saliency detection under the conditions of complex scenes,a recurrent residual network based on dense aggregated features is proposed.The whole network consists four modules: dense features extraction module applys ResNeXt101 net to produce a set of dense feature maps with different scales;all features aggregation module aims at the transmission and fusion between semantic information and spatial details from various layers;recurrent residual module is exploited to progressively improve the saliency map under the deep supervision mechanism.Experimental results on public test datasets show that the proposed algorithm provides more relieable results than the existing algorithms.(3)In order to deepen the complemantarity and fusion of periodic feature information and improve the accuracy of visual saliency detection,a triple residual learning network is proposed,which includes dense atrous spatial pyramid pooling module equipped with cross residual learning,top-to-down residual learning and bottom-to-up residual learning modules.Firstly,DenseNet161 is exploited as the basic framework for features extraction.Through the dense connection and atrous convolution,stagewise features with large receptive field are obtained from different dense blocks.Then,cross residual learning is employed to different stage dense features for refinement.Besides,residual learning is also embed into the adjacent features and saliency map prediction in bidirectional cascaded way.Finally,the result of the last stage becomes the visual saliency detection output.Experimental results show that residual learning enhances the exchange and fusion of information between features,bidirectional cascaded structure enhances continuity of features and the proposed algorithms has reliable capability for visual saliency detection.